A sparse attack method on skeleton-based human action recognition for intelligent metaverse application

【Author】 Dai, Cheng; Huang, Yinqin; Chien, Wei-Che

【Source】FUTURE GENERATION COMPUTER SYSTEMS-THE INTERNATIONAL JOURNAL OF ESCIENCE

【影响因子】7.307

【Abstract】Graph convolutional networks (GCNs) based skeleton human action recognition has been one of the hottest techniques in intelligent Metaverse system applications in recent years. However, the robustness of GCN model based skeleton action recognition, has not been widely discussed due to the complex spatio-temporal nature. Base on solving this problem, we propose a sparse attack method with dynamic attention which can generate adversarial samples with lower deviation under the condition of keeping attack success rate in this paper. Firstly, we develop spatial temporal consistency loss which can not only preserve spatial integrity between reference sample and adversarial sample but also keep temporal coherence between consecutive frames across adversarial sample. Then, we develop an interaction-based perturbation contribution analyze method to discard unnecessary perturbations. Besides that, we design a dynamic attention approach to tilt the perturbations towards some joints that have higher dynamics which can decrease overall deviations. Finally, we conduct extensive experiments to evaluate our proposal and the experimental results show that our method contributes to discover potential causes of model fragility and provides material in adversarial training which will increase the robustness of GCN based skeleton human action recognition models.(c) 2022 Published by Elsevier B.V.

【Keywords】Graph convolutional networks; Skeleton human action recognition; Intelligent metaverse system applications; Spatial temporal; Sparse attack

【发表时间】2023 JUN

【收录时间】2023-03-08

【文献类型】

【主题类别】

--